Hugging Face enables Claude Code and Codex to directly participate in RL training.

P2P AI inference network Antseed has officially launched, with its non-profit operator Antseed Foundation disclosing that it has completed a $2.4 million token round financing. Spark Capital led the round, with participation from Collider, DCG, North Island Ventures, Reciprocal Ventures, Relay Capital, and Venice.ai. The project aims to adapt BitTorrent’s peer-to-peer model for AI inference, enabling different providers to directly sell models and agent services. After users run Antseed’s local router on their computers, they can connect tools like Claude Code, Codex, and OpenCode to a single address. Each inference call is routed by the router based on price, speed, reputation, and privacy criteria, with fees settled directly via USDC on Base. Providers can run their own local models, fine-tuned models, or other AI services, and set their own pricing. Its key difference from OpenRouter is the absence of a unified forwarding platform in the middle layer: buyers connect directly to providers, while Antseed handles discovery, routing, reputation, and payment protocols. Regular providers can still view the prompts they process; users also have the option to select TEE-verified providers, where requests run in a trusted execution environment (hardware-isolated from request content). Antseed is also refining verification mechanisms including model fingerprinting, signed responses, and auditing to prevent providers from passing off cheaper models as more expensive ones. The code is open-source and the network is operational, though compatibility issues with tools like Codex and Claude Desktop are still being addressed on GitHub. It is best described as a rapidly iterating "decentralized OpenRouter".

The Federal Reserve plans to overhaul the U.S. banking regulatory system, replacing the current model where regional Federal Reserve bank presidents oversee bank examinations with a new framework that holds Washington headquarters more clearly accountable. The Fed’s top banking regulatory official unveiled the reform plan on Tuesday. Michelle Bowman, the Fed’s vice chair for supervision, said the reform will establish five new geographic banking regulatory regions, each led by a "regional administrator". In remarks prepared for a St. Louis Fed conference, Bowman noted the current structure "undermines the critical link between responsibility and accountability". "The Fed’s supervisory functions will be restructured to build a culture that emphasizes accountability and clear decision-making authority," Bowman stated. Under the new regulatory setup, the five regional administrators will oversee all supervisory activities in their respective regions, though hands-on examination work will still be performed by staff at the regional Federal Reserve banks. Bowman also announced that the Fed will consider adjusting asset size thresholds later this year—thresholds that determine at what scale banks become subject to stricter regulatory rules.

Lido stated that following last week's infrastructure investigation, MetaMask Staking has essentially completed the exit process for its Lido validators, with no unexpected incidents occurring throughout the process. The exited ETH will continue to re-enter the Lido protocol over the coming weeks. Thanks to the operator's timely disclosure of the issue and rapid response measures, the impact on stETH holders has been limited. Lido noted that the final batch of validators is currently wrapping up their exits, with stETH's APR at 2.23%—matching or exceeding levels seen since August. stETH's returns come from over 600 independent node operators across its curated staking module and community staking module, so the preventive exit of a single large operator has a limited effect on overall returns. MetaMask Staking will release a full post-mortem report in the coming weeks.

Michael Burry, the real-life inspiration for *The Big Short*, today disclosed via his Substack account that he adjusted his portfolio for tax-loss harvesting: he sold Lululemon and shifted the position to Deckers, sold Fannie Mae to buy Freddie Mac; converted his Sprouts Farmers Market holdings into 2028 call options, swapped Zoetis stock for 2029 call options, increased his holdings of JD.com’s 2029 call options and BYD’s common stock, and added new 2029 put options on MetLife. Additionally, Burry is still waiting for Alibaba to fall further, while his Fiserv position remains unchanged. Burry noted that the fourth quarter "will bring additional pressure to stocks that have already declined sharply". Note: Tax-loss harvesting refers to the practice of selling loss-making assets to realize paper losses, which can offset capital gains generated from other investments to reduce tax liability.

Beating AI Express: AI Agent observability platform Laminar has launched flow-1, a model purpose-built to inspect Agent execution traces. The model reads model calls, tool calls, and return results to pinpoint where an Agent errs and the root cause of the mistake. flow-1 operates on Laminar’s Signals agent, capable of searching the full Agent execution trace and drilling down into specific steps as needed. For training, Laminar first used synthetic survey data for supervised fine-tuning, then further trained the model on tool calls and complex trace analysis via reinforcement learning. In Laminar’s self-built benchmark of 523 challenging traces, flow-1 posted an error detection F1 score of 0.835, outperforming GPT-6-sol’s 0.816. GPT-6-sol has a higher recall rate, catching more actual errors, while flow-1 delivers higher precision with fewer false positives. For traces under 100,000 LLM tokens, flow-1 costs an average of ~$0.0011 per analysis, compared to ~$0.026 for GPT-6-sol. Per Laminar’s calculations, $1 can analyze roughly 888 traces with flow-1 versus just 38 traces with GPT-6-sol, a cost difference of around 23 times. flow-1 is also approximately 25% cheaper than GPT-6-luna. All current results are from Laminar’s in-house benchmark, with no third-party verification available. flow-1’s training data is primarily synthetic workflows, around 48% of which are software engineering tasks. Some community members have questioned whether the model can maintain its performance when facing novel, unforeseen failures in real production environments.

Beating AI News: French AI firm Mistral has launched its new flagship model, Mistral Large 4, internally codenamed "Le Chonk". It adopts a Mixture of Experts (MoE) architecture, with a total of 1.05 trillion parameters, though only 49 billion are activated per inference. The model natively supports text and image input, and offers a maximum context length of 1 million tokens. Its API is now available for preview, while the model weights are scheduled to be released publicly on October 27. Mistral claims Large 4 is currently the highest-performing open-weight model developed in the U.S. and Europe. Its official DeepSWE v1.1 software engineering benchmark score stands at 62%, compared to GLM-5.3 (61%), DeepSeek V4 Pro (57%), and Qwen 3.8 Max (51%). For financial benchmarks, Finch scores 67%, on par with DeepSeek V4 Pro; satellite image target localization task DIOR-RSVG scores 73%, higher than GPT-6 Astra's 68%. However, these results are primarily from Mistral's self-testing, not based on a unified standardized testing framework. When Zhipu AI launched GLM-5.3, it announced a score of 66.9% on the same DeepSWE v1.1 benchmark, while Kimi K3 reached 67.5% – both outperforming Large 4's 62%. The gap may stem from different testing configurations and execution methods used by various teams. Large 4 is trained from scratch: Mistral used around 4,000 Nvidia Grace Blackwell GPUs in its own European data centers over approximately two months. Mistral also positions "European autonomy" as a key selling point, noting that the model – from training and API services to future independent deployment – can all be hosted on European infrastructure.